HiveMind banner

HiveMind: A Self-Hosted Team of Always-On AI Agents

Most agent frameworks hand you a workflow builder and wish you luck. HiveMind hands you colleagues.

HiveMind is a self-hosted system for running a team of persistent, named AI agents on Linux — written in C++23, MIT-licensed, currently at v0.9.5. You create teammates, message them in threads like people, and hand off real work. Each agent runs in its own rootless podman container with passwordless sudo, remembers everything across restarts, and only comes back when it needs your approval.

The repo is live at gitlab.com/leestripp/hivemind.

Chat, not config

There's no workflow editor. You talk to agents in DMs or group threads, and threads are the orchestration model — fan-in when a task needs more than one teammate. (The 1.x node-graph engine was retired in Phase 12. Good riddance.)

A computer of their own

Each agent works inside a rootless podman container as user hive with passwordless sudo. They can apt install whatever CLI tools the job needs. A shared company /workspace volume lets the team pass work around; a private /home/hive keeps each agent's own state separate.

Teamwork is the point

Agents discover, message, and delegate to each other through list_agents, message_agent, and create_agent tools. An agent on a remote rig can send_file finished work straight into the company workspace on another machine. A chief-of-staff agent can run the whole team. And when an agent produces an artifact, attach_file drops it into the thread as a normal chat attachment — images and video render inline.

Approvals stay with you

Flip require_approval on an agent and its commands park in an approval queue until you approve or deny. Autonomy where you want it, a brake pedal where you don't.

Distributed workers

Enroll any Linux machine as a hive-worker and assign agents to it — their containers, exec, and screen sessions run on that machine over a persistent, auto-reconnecting WebSocket channel. Enrollment is a one-time key:

hive-server enroll-worker <name>
hive-worker --server wss://<server> --key hmwk_... --name <name>

TLS runs through the nginx proxy; a direct ws://<server>:8423 channel works on a trusted LAN.

The pitch

If you've wanted AI agents that behave like a dev team instead of a prompt pipeline — persistent, file-sharing, approval-aware, and spread across every machine you own — this is worth a look. MIT license, C++23, Ubuntu 26.04, v0.9.5. Self-hosted means your agents, your machines, your data.

HiveMind on GitLab →